All Languages Matter: Swahili ↔ Sukuma Machine Translation

In 2024 we built Kamusi, a way to learn, look up, and explore Swahili with AI nearby, and even then, the point was never Swahili alone. We believe all languages matter, and that intelligence should live in the languages people actually speak and write, not only the ones that already crowd the internet.

Sukuma is spoken by millions, and it should not wait at the edge of the map while tools keep arriving for everyone else. So we opened small Swahili → Sukuma translators as a research preview, kept small on purpose so they can run on phones, everyday machines, and places where big cloud GPUs are scarce and the network is thin. Type Swahili below and you should see Sukuma (Kisukuma) come back.

Why two forward versions?

Not every device has the same room to spare, so we share two sizes for Swahili into Sukuma. Reach for Quality when you can afford a stronger model, or Lite when memory and speed have to win. Both are on Hugging Face (quality · lite).

The other way: Sukuma → Swahili

Translation is a loop, not a one-way street. We now also ship a complementary Sukuma → Swahili model, nileagi/nileagi-suk-mt-rev, trained on the same literary Swahili–Sukuma sentence table as the forward SKUs, with source and target swapped. Splits stay whole document groups (29,962 / 993 / 147) so neighbouring lines from one document do not leak into the test set.

On that held-out test set the reverse SKU reaches chrF2 39.5 and BLEU 11.7. Forward quality on the same split (opposite direction) sits a little higher (chrF2 42.6 / BLEU 20.0). A chrF2 near 40 usually means meaning is often carried, but the string is not a clean match to the reference—still a research preview, not a production translator.

Sukuma input should use the literary macron orthography (ā ē ī ō ū). Macron-free Sukuma looks like domain shift. Swahili output is standard Latin Swahili. Short formulaic lines often land in readable Swahili; longer literary sentences are less stable—names and rare stems get approximated. Prefer chrF2 when you score new text, and do not treat outputs as a substitute for a human translator.

Try both directions in the lab preview; the blog box above stays Swahili → Sukuma for a quick first look.

What to expect

This remains a research preview on literary, read-aloud register. Casual chat, legal wording, medical phrasing, and news are out of scope. English is not a source or target of these SKUs—if you start from English, translate to Swahili first (or use the lab's Transformers script path), then call Sukuma. Treat the demos as living previews, and the open weights as a starting point for anyone who wants Sukuma in the loop rather than a finished product for every domain.

Applications

The point of this preview is not a production translator for every domain. It is a way to put Sukuma in the loop on devices people already have—both into Sukuma and back into Swahili.

• Learning and lookup

  Give Kamusi-style tools a Swahili↔Sukuma step so learners see Kisukuma beside the Swahili they already know, and can check a Sukuma line in Swahili.

• Community notices and local content

  Draft Sukuma from a Swahili source for radio, church, school, or village notices—or draft Swahili from a Sukuma source—then have a speaker edit before it goes out.

• On-device translation

  Run the smaller forward checkpoint where the network is thin, and the reverse SKU when you need Swahili out of literary Sukuma, without waiting for a cloud GPU that never learned the language.

For builders

For Swahili → Sukuma, force the Sukuma target tag (suk_Latn) with Swahili as source (swh_Latn). For Sukuma → Swahili, set source suk_Latn and force BOS swh_Latn—omitting either side will not match the reported scores. The lab try box already does the right tags for you. Methods, scores, and limits for forward are in the technical report below; reverse details live with the reverse Hub repo.

Collaboration

If you are extending this work, evaluating on new data, or building with Sukuma in your community, write to us at hi@nileagi.com. The weights are released under CC BY-NC-SA 4.0; commercial use needs a written agreement with NileAGI.